Details
-
Improvement
-
Resolution: Incomplete
-
Major
-
None
-
4.6.4, 4.6.3, 5.0.0, 5.5.0
-
CentOS 7.3
CPU E5-2630 v3 (32 vCPU)
64GB RAM
Samsung SM863 SATA SSD
Description
Test scenario:
- 4 nodes
- 1 bucket, 1 replica, full eviction
- 1B items, ~10% resident ratio
- 40 async client threads, random read & modify operations (uniform distribution)
Currently, the maximum throughput that we can achieve during this workload is ~35K total ops/sec (17K gets/sec and 17K sets/sec). That is about 4K disk fetches/sec (ep_bg_fetched) and 9K disk updates/sec (ep_ops_update) per node in one of the latest 5.1.0 builds.
Actual disk activity is significantly higher - about 30K disk reads per second per node (or 300 MB/sec). See rps and rbps graphs for details. Writes seem more efficient because of batching.
Keep in mind that 35K ops/sec is not sustainable throughput. There were almost 10M items in the disk write queue at the end of the test.
Graphs with compaction enabled and disabled: